loading
Loading.dokan · open source
dokan is a deterministic job runtime for AI agents. Your agent scripts the repeatable work and runs it over MCP; dokan executes it in clean containers with no model in the loop, so it burns zero tokens.
Open source, Apache-2.0 · public beta · zero LLM inside
Works with
Claude Code · Cursor · Windsurf · any MCP client
Deterministic code, not a model in the loop. The expensive intelligence stays in your agent; dokan just runs the mechanical 80% cheaply, and the platform never spends a token to execute your workflows.
Your agent uploads scripts, wires them, triggers runs, and reads logs over MCP. You don't touch a dashboard.
One job, one clean container: network off by default, per-job CPU and memory caps, timeouts, retries. Identical inputs hit a content-addressed cache instead of recomputing, and every run carries a tamper-evident receipt.
Cron schedules and inbound webhooks (Stripe, Calendly, GitHub-ready), so your agent's workflows fire on a schedule or on an event, not only on demand.
Your agent talks to the dokan daemon over MCP. dokan runs each job in its own container, network disabled by default, captures the output, and posts the result back to your agent. State lives in Postgres; the agent reads IDs and paginated logs, not payload dumps, so it stays inside its context budget.
The agent scripts the repeatable 80%, dokan runs it cheaply and reliably, and nobody clicks through a UI.
Public beta. Open source (Apache-2.0). Built and run in production by tsukumo, the studio behind it.
When we ship or learn something real about running deterministic work for agents. No filler, no fixed schedule.
dokan is one of four open-source tools tsukumo builds and runs in production (trovex, wrai.th, yoru, dokan). Running agent fleets on real work and want this set up with your team? tsukumo.ch